AI Agents Can Write Passing Tests That Prove Nothing — Here's the Fix

A software developer observed that when AI agents write both code and its accompanying tests, the tests lose their value as independent verification since they simply reflect the same model's assumptions twice. Unlike human-written tests, which carry a second perspective, agent-generated tests are better described as claims rather than evidence. The developer argues the cheapest way to validate such tests is to watch them execute, not just read the diff or assertions. A tool called twd-cli addresses this by recording actual test runs as short video clips — one per test — paced to be humanly watchable rather than a millisecond blur. Because the recordings are generated by the live test execution, they stay in sync with the codebase and cannot drift the way manually recorded demos do.